DeepScaleR: Surpassing O1-Preview with a 1.5B Model by Scaling RL

Re:DeepScaleR: Surpassing O1-Preview with a 1.5B Model by Sc... by jazzyjackson on Tuesday February 11, 2025 @05:21PM (#43019114)

Nah, it takes a mastermind to be creative. The tiny models IME are good at following explicit instructions for data transformations (conforming to a JSON schema is especially helpful) but I wouldn't expect one to solve a problem I haven't already solved myself like how I use the .6T full size models.

[ Parent | on HN ]

Re:DeepScaleR: Surpassing O1-Preview with a 1.5B Model by Sc... by EE84M3i on Tuesday February 11, 2025 @05:25PM (#43019157)

I thought conforming to a JSON schema was a solved problem for even the most garbage models by restricting the tokens it can generate?

[ Parent | on HN ]

Re:DeepScaleR: Surpassing O1-Preview with a 1.5B Model by Sc... by jazzyjackson on Tuesday February 11, 2025 @08:35PM (#43020763)

Yes that's what I mean, with a well described task and a JSON schema, the tiny models are suitable (though still hallucinatory - even llama3 7B was making up phone numbers when I tasked it with extracting contact details from email signature blocks)

I was pushing back against the parent's notion that tiny models would serve general purpose needs.

[ Parent | on HN ]